m-mohr
commented
1 month ago Disclaimer: I'm not sure whether there's a best practice for versioning ML models.
We could version these models in a simple way I think, just incrementing numbers (e.g. v1, 2, 3 etc). We probably don't need minor (0.x.0) or patch versions (0.0.x). Right now there are no version numbers associated with the release: https://github.com/fieldsoftheworld/ftw-baselines/releases That makes it difficult in provenance to track the models. It also doesn't allow to simply reuse the (non-existing) version number to release on pip, so it's hard to release there in a way that there's a direct relation between model and software releases. The missing versioning also results in the question: What would be the next release? Overrriding the old one?
We could then also release the software to pip alongside the model releases, but we should then break out software that's not used for the model creation, see #28 . This would allow to iterate on tools faster than the models evolve.
Not sure I understand the releases made so far correctly, but it seems like we should rename the releases to: v0 -> https://github.com/fieldsoftheworld/ftw-baselines/releases/tag/model v1 -> https://github.com/fieldsoftheworld/ftw-baselines/releases/tag/Pretrained-Models
The config files should probably also be tagged instead of being duplicated each time?! https://github.com/fieldsoftheworld/ftw-baselines/tree/main/configs (although it seems that you are using 1.x.x version numbers there)
Thoughts?
We should define a way to version this, I can't find any version numbers.
Also, what's the release process? This should be documented so that everyone knows how it works.
Happy to propose something... just quickly opening this issue so that I don't forget.